Can voluntary frameworks between governments and private AI firms adequately govern military uses of AI? Analyse.

Military AI is now built mostly by private firms. In May 2026 the U.S. Department of War signed agreements with eight companies, including OpenAI, Google, Nvidia and Microsoft, to run their tools on its highest-tier classified networks (Impact Levels 6 and 7) [1]. A White House meeting with AI CEOs followed in September 2026 [2], and it ended in a voluntary accord. Such pledges help, but they cannot govern military AI on their own.

What voluntary frameworks offer

• Speed: a pledge can be revised as fast as the models change. Treaty talks under the CCW Group of Governmental Experts take years [3].
• Industry expertise: the firms that build the models understand their technical failure points best.
• Stop-gap where consensus is missing: major powers, including the U.S. and Russia, voted against the 2025 UNGA draft on autonomous weapons [4].

Why they fall short

• No enforcement: a firm that breaks a pledge faces no legal penalty. Yet its systems now sit inside the most sensitive military operations [1].
• Accountability gap: States still disagree on what "meaningful human control" means [5]. A corporate pledge cannot settle who answers for an unlawful strike. As one First Committee speaker put it, an algorithm must not be in full control of decisions involving killing [3].
• Concentration of power: cloud, chip and model firms also dominate civilian markets. Once they run daily operations they are hard to replace (vendor lock-in), which gives them leverage over the state.
• No independent review: the UN advisory body asks for red lines, such as AI selecting and engaging targets on its own, and for legal and technical reviews before deployment [6]. Pledges provide neither.
• Weakens multilateral norms: when one State relies on its own pledges, global rule-making slows. The UN Secretary-General wants such weapons banned under international law [7].

What adequate governance needs

• National law that requires a pre-deployment legal review and names a human legally responsible for each system [6].
• Multilateral rules under the CCW. India voted against the 2023 UNGA draft because it prefers the CCW track [3]. It can now help shape the definition of human control instead of only blocking texts [5].

In short, voluntary frameworks are a useful first layer of governance, but they cannot be the whole structure. Rapid AI adoption should be paired with binding national reviews and multilateral standards. That would keep international humanitarian law and human accountability at the centre of decisions to use force.